Research On Abnormal Transaction Data Analysis Of Blockchain Based On Machine Learning

At present,the most effective detection method for Ponzi schemes is a combination of machine learning methods,but the selected model is still slightly insufficient.The method of manual detection requires the source code of a Ponzi scheme,which cannot handle the thousands of new smart contracts and a large number of non-open source smart contracts deployed every day;opcode frequency detection can lead to potential Ponzi scheme omissions;and account feature detection will simplify the data set,which may lead to model overfitting and low detection performance.In order to solve some of the existing problems and further improve the performance of model detection,this thesis adopts the extreme gradient boosting（e Xtreme Gradient Boosting,XGBoost）model with the best comprehensive effect combined with the sparse perception algorithm,which automatically learns the splitting direction to obtain the optimal objective function by automatically learning the splitting direction for the missing values in the sample data,and finally improves the performance of the model performance by using the improved greedy algorithm to improve the performance of the Ponzi scheme detection in the smart contract.Construct a ROC（Receiver Operating Characteristic curve,ROC）curve comparison chart.The idea of generating new sample data by generating adversarial networks is proposed to reduce overfitting.The main contents and innovations of this article are as follows:（1）Combined with the improved greedy algorithm,the detection performance of the existing XGBoost model is improved.For anomalous transaction data detection in Ethernet smart contracts,this thesis compares the performance of three commonly used machine learning models: Support Vector Machines（SVM）,Random Forest（RF）,and XGBoost.Compare the advantages and disadvantages of each model detection,and improve the detection performance of the best and most practical XGBoost model combined with the improved greedy algorithm.The model constructed in this thesis can effectively improve the model training speed,reduce the complexity of the model,and effectively improve the Call,F-score and AUC（Area Under Curve,AUC）values of the model.（2）Based on the Convolutional Neural Network model,the probability that a smart contract in the blockchain is a Ponzi scheme is detected.This thesis can input text data into the image processing model through the CNN model,which is convenient for generating new sample data with the generative adversarial network model in the future,and fundamentally solves the problem of insufficient sample size.In addition,after adding K-fold verification to the CNN model in this thesis,in addition to improving the performance,the probability identification and analysis of the characteristics of Ponzi scheme in each account are analyzed,and the probability heat map of Ponzi scheme is detected,which more effectively improves the accuracy and credibility of the model.The user simply enters the account characteristics in the transaction information into the model to identify whether the transaction has an abnormal transaction by a known rate,and if the alert may be a Ponzi scheme,further confirm the mark with other account characteristics.